Enable constant propagation for more math functions

Thu Aug 27 12:58:49 PDT 2015

I found another issue: acos(3.0) returns NaN, and thus isn't constant
folded. Trying again...

-erik

On Thu, Aug 27, 2015 at 3:10 PM, Erik Schnetter <schnetter at gmail.com> wrote:

> Vedant
>
> Thanks for the pointer. I think that "CHECK: ret" statements are missing.
>
> -erik
>
> On Thu, Aug 27, 2015 at 2:38 PM, Vedant Kumar <vsk at apple.com> wrote:
>
>> check-all runs fine on my machine too. From the logs, I see:
>>
>> >> <stdin>:372:7: error: CHECK-NOT: string occurred!
>> >>  %1 = call double @acos(double 3.000000e+00)
>> >>       ^
>> >>
>> /home/llvmbb/llvm-build-dir/clang-x86_64-debian-fast/llvm.src/test/Transforms/ConstProp/calls.ll:421:14:
>> note: CHECK-NOT: pattern specified here
>> >> ; CHECK-NOT: call
>>
>>
>> However, line 421 of calls.ll tests is:
>>
>>     "%0 = call float @tanhf(float 3.000000e+00)",
>>
>> and the output string FileCheck matches against is:
>>
>>     "%1 = call double @acos(double 3.000000e+00)" (from line 430).
>>
>> So I suspect that the code is fine but the test case is flaky. I will
>> keep looking at this, but I thought I'd shoot this email back at you now in
>> case it helps :).
>>
>> vedant
>>
>>
>> > On Aug 27, 2015, at 11:14 AM, Erik Schnetter <schnetter at gmail.com>
>> wrote:
>> >
>> > This commit led to test suite failures (see e.g. <
>> http://lab.llvm.org:8011/builders/clang-x86_64-debian-fast/builds/30501>),
>> so I reverted it.
>> >
>> > Locally, everything is fine if I run "make check". Do you have pointers
>> for debugging this?
>> >
>> > -erik
>> >
>> > On Thu, Aug 27, 2015 at 12:41 PM, Erik Schnetter <schnetter at gmail.com>
>> wrote:
>> > Committed as 246158.
>> >
>> > -erik
>> >
>> > On Wed, Aug 26, 2015 at 1:51 PM, Vedant Kumar <vsk at apple.com> wrote:
>> > LGTM, this is ready to land.
>> >
>> > It turns out that we _can_ combine these tests. I'll take care of it in
>> a followup patch since existing tests need to be changed.
>> >
>> > > >> +define float @test_atan2f() nounwind uwtable ssp {
>> > > >> +entry:
>> > > >> +; CHECK-LABEL: @test_atan2f(
>> > > >> +; CHECK-NOT: call
>> > > >> +  %0 = call float @atan2f(float 3.000000e+00, float 4.000000e+00)
>> > > >> +  ret float %0
>> > > >> +}
>> > > >>
>> > > >> +; Shouldn't fold because of -fno-builtin
>> > > >> +define float @atan2f_() nounwind uwtable ssp {
>> > > >> +; FNOBUILTIN-LABEL: @atan2f_(
>> > > >> +; FNOBUILTIN: %1 = call float @atan2f(float 3.000000e+00, float
>> 4.000000e+00)
>> > > >> +  %1 = call float @atan2f(float 3.000000e+00, float 4.000000e+00)
>> > > >> +  ret float %1
>> > > >> +}
>> > >
>> > > I've noticed many pairs of test functions like the one above. Can
>> this pair be combined into one function? E.g:
>> > >
>> > > >> +define float @atan2f_() nounwind uwtable ssp {
>> > > >> +; CHECK-LABEL: @_atan2f(
>> > > >> +; FNOBUILTIN-LABEL: @atan2f_(
>> > > >> +; CHECK-NOT: call
>> > > >> +; FNOBUILTIN: %1 = call float @atan2f(float 3.000000e+00, float
>> 4.000000e+00)
>> > > >> +  %1 = call float @atan2f(float 3.000000e+00, float 4.000000e+00)
>> > > >> +  ret float %1
>> > > >> +}
>> > >
>> > > I don't know, I'm not too familiar with the test system. These are
>> two different tests: in one, the call should be eliminated
>> (constant-folded), in the other case, it should not. Do you have a pointer
>> to documentation or other examples that cover this detail?
>> >
>> > Your ">> +; CHECK-NOT: call" lines check that there is "NOT" a "call",
>> i.e the call is folded away. When "-fno-builtin" is passed to the compiler
>> however, we want the call preserved. FileCheck lets you specify different
>> check prefixes for different usage scenarios.
>> >
>> > Googling "llvm-lit" or "llvm FileCheck" should turn something up.
>> >
>> > thanks!
>> > vedant
>> >
>> > >
>> > > If so, it would be nice to reduce the test case size. Apart from
>> that, please run clang-format -style=llvm on your changes. Everything else
>> looks good.
>> > >
>> > > I attach the updated (re-formatted) patch.
>> > >
>> > > -erik
>> > >
>> > > vedant
>> > >
>> > > On Aug 25, 2015, at 1:14 PM, Erik Schnetter via llvm-commits <
>> llvm-commits at lists.llvm.org> wrote:
>> > >
>> > > > It seems that constant propagation for single precision math
>> functions (such as tanf) is already working, but was not enabled. This
>> patch enables these for many single-precision functions, and adds
>> respective test cases.
>> > > >
>> > > > List of newly handled functions: acosf asinf atanf atan2f ceilf
>> coshf expf exp2f fabsf floorf fmodf logf log10f powf sinhf tanf tanhf
>> > > >
>> > > > -erik
>> > > >
>> > > > --
>> > > > Erik Schnetter <schnetter at gmail.com>
>> http://www.perimeterinstitute.ca/personal/eschnetter/
>> > > > <constprop.diff>_______________________________________________
>> > > > llvm-commits mailing list
>> > > > llvm-commits at lists.llvm.org
>> > > >
>> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.llvm.org_cgi-2Dbin_mailman_listinfo_llvm-2Dcommits&d=BQIGaQ&c=eEvniauFctOgLOKGJOplqw&r=nRkKAYlatRp-Wgk-tAtmRg&m=zIeTq7C24I26_ny62rhiVNnPUKhiL-ttf5E5SdvSYfM&s=hnzB8DZ-EJj3yM7S-bYX5mwxBMka0n18KAACB3fEHcQ&e=
>> > >
>> > >
>> > >
>> > >
>> > > --
>> > > Erik Schnetter <schnetter at gmail.com>
>> http://www.perimeterinstitute.ca/personal/eschnetter/
>> > > <constprop.diff>_______________________________________________
>> > > llvm-commits mailing list
>> > > llvm-commits at lists.llvm.org
>> > >
>> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.llvm.org_cgi-2Dbin_mailman_listinfo_llvm-2Dcommits&d=BQIGaQ&c=eEvniauFctOgLOKGJOplqw&r=nRkKAYlatRp-Wgk-tAtmRg&m=pdx8WaNzTKH4XL-GadzDAXtcW_0Dy5TSuEq-r6kr_Cw&s=p1rycXaVwpkLIJ6ivjsTu1JL_CRTjkncrT0wSb31Qy8&e=
>> >
>> >
>> >
>> >
>> > --
>> > Erik Schnetter <schnetter at gmail.com>
>> http://www.perimeterinstitute.ca/personal/eschnetter/
>> >
>> >
>> >
>> > --
>> > Erik Schnetter <schnetter at gmail.com>
>> http://www.perimeterinstitute.ca/personal/eschnetter/
>>
>>
>
>
> --
> Erik Schnetter <schnetter at gmail.com>
> http://www.perimeterinstitute.ca/personal/eschnetter/
>

-- 
Erik Schnetter <schnetter at gmail.com>
http://www.perimeterinstitute.ca/personal/eschnetter/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150827/67425ac0/attachment.html>